RETRIEVAL-AUGMENTED GENERATION: RAG, Vector Databases and Knowledge-Driven AI : Foundations, Architectures, Evaluation and Production Systems for Grounded Language Models
English | September 21, 2026 | ASIN: B0HKJCJ8YJ | 327 pages | Epub | 5.40 MB
In the space of a few years, large language models moved from laboratory curiosities to instruments that draft our contracts, summarise our clinical notes and answer our customers. Yet anyone who has deployed such a model in a serious setting has met the same wall: the model is articulate about everything and accountable for nothing. It will describe a refund policy that does not exist, cite a statute that was never enacted, and do so in prose of impeccable confidence. The model is not lying; it has no notion of a fact to lie about. It has learned a probability distribution over text, and a probability distribution has no bibliography.Retrieval-augmented generation is the engineering answer to that predicament. Instead of asking a network to remember the world inside its weights, we let it look things up. A retriever finds evidence, a generator reasons over that evidence, and the answer arrives with a trail back to its sources. The idea is simple enough to state in one sentence and deep enough to occupy an entire research literature, because everything interesting lies in the details: how text is cut into passages, how meaning is turned into geometry, how a billion vectors are searched in milliseconds, how conflicting evidence is reconciled, and how we know whether any of it worked.This book is my attempt to set those details down in one place, in the order in which a reader actually needs them.Who this book is forThe book is written for three audiences that I have taught and worked with, and it tries to serve all three without diluting any of them.• Senior undergraduate and postgraduate students in computer science, data science and information retrieval who need a rigorous course text rather than a collection of blog posts. Each chapter ends with a summary and graded exercises suitable for coursework.• Practising engineers and architects who must ship a retrieval system that is fast, cheap, auditable and secure, and who need the reasoning behind the defaults, not only the defaults.• Researchers and technical managers who want a map of the field: what is settled, what is contested, and where the open problems are.The mathematical prerequisites are modest and mostly confined to linear algebra and elementary probability; Appendix A collects everything the main text assumes. The programming examples are in Python, which is where the ecosystem lives, but the ideas are language-agnostic.How this book is organisedPart I builds the foundations: why parametric memory fails, how text becomes vectors, what an embedding space really is, how vector databases store and index those vectors, and the algorithms that make approximate nearest-neighbour search tractable. Part II constructs a production pipeline end to end: ingestion, parsing, chunking, indexing, hybrid retrieval, query transformation, re-ranking, context assembly and grounded generation. Part III treats the advanced systems that define the current frontier: self-correcting and agentic architectures, knowledge graphs, multimodal and long-context retrieval, evaluation, optimisation and security. Part IV puts everything to work through four complete, buildable projects and a set of deployment case studies.Chapters are meant to be read in sequence the first time and consulted out of order thereafter. Cross-references are given by section number so that the book remains usable as a reference.
Links are Interchangeable - Single Extraction
